AMD gives 7000-series Threadrippers a frequency bump with Epyc core counts AMD has launched 7,000-series Ryzen Threadripper processors, bringing its 5nm Zen 4 cores to high-end desktops and workstations. As with previous generation Threadripper parts, the chips borrow heavily from their Epyc datacenter counterparts that we reviewed last November, and feature the familiar combinations of compute and I …
Video encoding, 3D rendering, code build/analysis of multi-file projects where you could allocate one core per file... Sure, some of these can now be offloaded to a GPU if you've got a suitable one available as well, provided your application supplier has gone down that route, but there's still a lot of stuff that remains firmly wedded to CPU processing which could therefore take advantage of more cores.
I'm hoping to get one for a work. I do my design evaluations using Ansys Mechanical workbench and Fluent. Most times we send the job to a solver but the model prep is done on our workstations. Meshing large models can take a while and scales very well with cores. Not too hard to make a case for an upgrade and its nice when you can solve smaller jobs locally. Good single thread performance for poorly threaded loads like Solidworks is a great bonus. I know I want one.
> Meshing large models can take a while and scales very well with cores.
How many cores does your current system have and are you max’ing them (and the memory I/O channels) out when meshing.
Just interested, as to whether the problem really does scale linearly or you reach a peak performance point because other factors limit the extent to which you can parallel process. Ie. Going from 8 to 16 cores is a massive improvement, 32 a slight improvement because you only really needed 21.
My current workstation is a 5 year old six core (Lenovo P520 Xeon Intel W-2135) that overdue for an upgrade. It's fine on small simulations since the setup takes a lot more time than solving (checking off various buttons and selecting materials). It seems to scale well for what it is. We have a few dual CPU Xeon's (P720's), they have more cores but run at lower clock speeds and don't scale very well. In general the lower core counts with better clock speeds was better. The dual cpu workstations are about 3 years old.
That's what I have been using for most daily tasks.
Last year, Instead of everyone in the department getting a new workstation we decided on a shared resource that we could send our jobs to. For that we got a 64 core Threadripper Pro 5995WX with 512GB of memory set up as a solver. That also saves having to outfit every workstation with a huge amount of memory. On my P520 128GB works just fine as long I I have a solver that I can send large jobs to. Large jobs can easily eat up 250+ GB
When we send medium sized jobs to it, it scales well up until about 20-24 or so cores. A job that took an hour to solve on my older P520 is completed in less than 20 minutes. On large jobs that were taking from 8 to 12 hours to solve on the old systems the 5995WX is typically 3 to 4 times faster and scales well to about 32 to 36 cores. Most of this was informal testing and observations that when re-running a simulation with a changed core count there was a point of diminishing returns.
That corresponded with what I had read before I bought the system. I got the 5995WX for the larger L3 cache (compared with the 32 core version) since that was reported as a 12% improvement at 32 cores compared with the 32 core 5975WX. Jobs that were taking 8 hours are done in 2.5 hours. It lets us see the results and run a couple of iterations in the same workday, my coworkers were super happy with the setup. A single system was plenty for four engineers since the demand is not constant, but when you are in that phase of work you need it. It's managed to cue the jobs up so no more than two jobs run at a time. In the past year we only had one day where that was an issue.
We are getting a second solver to pool resources and adding users to the group.
The short story is that the system scales well to 32 or so cores and me belief is that bandwidth to system memory limits the performance after that. I'm curious to see what the new TR systems with DDR5 will do. From what I have read the newer Xeon's with DDR5 do quite well on CFD and the new 7000 Threadrippers will be a significant improvement. I have been holding off on upgrading my own system waiting for these to be released.
I was OK with my old desktop computer for daily tasks until I decided to setup a simulation on the Threadripper using a remote session. That made my P520 seem really, really slow (because it is).
My goal is to upgrade our local workstations with 24 or so core systems to get better single thread speed for general CAD and reduce setup time and still used pooled solvers to send the jobs to. GPU solvers have been looked at but not tried. A concern is most of our jobs are run using double precision and that does not seem to be a strong point of most GPU solvers. The other is the size of the models require lots of ram or a change in the solve type. The cost to get to a GPU solver that was competitive was considerably higher. I would love to be proven wrong on this and my current TR system has six open GPU bays. Our IT group is supposed to be getting us a GPU demo board to evaluate. The white papers I read were using four GPU's to get some amazing results but the cost was kind of crazy. The double precision thing seems to be a sticking point.
We have one person sending jobs to a cloud based solver. That seems to work OK but there is concern over the extra overhead and sending info outside our local network (he is one of the reasons we are getting another solver). It would be nice to have unlimited budgets and time to try various combos.
Mostly I just want a faster PC!
Really useful feedback.
In reading your workflow and use of shared resources, I am reminded of the ideas behind Apollo Domain distributed computing environment, where the OS sends the jobs to appropriate resource. Looking at modern OS’s it seems such useful capabilities are a long way down the to-do list (that’s if they are even on it); unlike UI frippery…
Aside: I assume you are using 10 Gbps LAN.
The work I've done in Comsol has similar requirements. RAM speed and quantity is king, processor very much secondary.
The moment you run out of RAM and start swapping to disk the CPU choice is irrelevant!
That said, short of some very fat server boards, there is little better than the thread ripper line up.
My old work machine for this task is getting long in the tooth. Maybe an upgrade is in order!?
Threadripper is not aimed at average PC users or to run typical desktop applications.
Threadripper supports ECC memory and has more PCIe lanes than what Intel or AMD Zen offers.
More cores are helpful with (nested) virtualisation.
With Threadripper rig you can have a test/dev environment at your own desktop for cheaper than using actual Xeon / EPYC servers. YMMV.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
